News vs. Noise: AI Deal Raises New Credit Questions

The yen carry trade is thinning, and the risk it carries has spread beyond the currency market

For more than two decades, investors have borrowed cheap yen and converted it into dollars to buy higher-yielding assets — a trade that works as long as Japan’s interest rates stay well below everyone else’s. That gap is narrowing. Japan’s policy rate now sits at 1%, against the Federal Reserve’s target range of 3.5% to 3.75% [1], and as of today’s close, Monday, August 17, the 10-year Treasury yield stood near 4.69% and the 30-year had climbed to 5.31%, its highest level since June 2007, while the dollar was buying about 159.5 yen [2]. Spot gold traded near $4,395 an ounce, roughly 21% below its all-time high of $5,589, set on January 28 [3].

A claim circulated last week that Japan selling its Treasury holdings is the main force pushing the 10-year toward 4.75%. Federal Reserve data show no foreign-official draw on the Fed’s FIMA repo facility, the tool that lets a central bank like Japan’s swap Treasuries for dollars instead of selling them outright, through the week ending August 5 [4]. That points to domestic forces — federal borrowing, inflation, and questions about Fed credibility — as the bigger driver of the move, not Japanese selling [1]. It also draws a sharper line around where the real risk sits: the facility only protects government-to-government transactions. It does nothing to stop a private hedge fund from selling U.S. stocks or emerging-market bonds to repay a yen loan if the currency strengthens quickly, and that private side of the trade is far larger and far less visible than the government side [1].

We are watching the dollar-yen exchange rate as the signal that matters most here, not any single day’s Treasury move. We are staying diversified across currencies and geographies rather than chasing the higher yields this trade has funded for years, because the goal is not calling the exact day it unwinds — it is not being caught holding the leveraged side of it when it does.

A softer midweek inflation reading lifted rate-cut hopes — then Friday’s consumer data cut against it

On Wednesday, August 12, July’s Consumer Price Index (CPI) came in close to economists’ forecast, and by the next day, traders had pushed the odds the Federal Reserve holds rates steady in September to about 65%, up from roughly 50% just days earlier, with the semiconductor index rising 2.5% and oil prices easing below $90 a barrel on the relief [5]. That relief looked fragile even at the time: Brent crude had spiked to $89.46 just a day before, on tensions over shipping through the Strait of Hormuz, a reminder that a single inflation print does not capture an oil move that happens after the data is collected [6].

Two days later, on Friday, August 14, that fragility showed up directly in the data. Retail sales fell 0.6% from June, the steepest monthly drop since May 2025 and well below the 0.1% gain economists expected [7]. Released that same morning, the University of Michigan’s preliminary August consumer sentiment reading dropped to 51.0 from July’s 55.2, snapping two straight months of improvement, driven mainly by how people feel about business conditions ahead rather than their own household finances [7]. Both readings trace back to the same source: gasoline prices climbing again as the U.S.-Iran standoff over Hormuz disrupts oil supply, squeezing household budgets and confidence at the same time [7].

We are watching whether Friday’s numbers mark a one-off soft patch or the start of a genuine pullback in consumer spending, which drives roughly two-thirds of U.S. economic output. A Fed that is not offering clear forward guidance means rate expectations can swing meaningfully within a single week, which is exactly why we stay diversified across the yield curve rather than positioning client portfolios for one outcome to resolve quickly.

NVIDIA’s $500 billion AI financing plan is turning the AI buildout into a credit story

On Monday, August 10, NVIDIA signed nonbinding agreements with six Wall Street financial firms — private-credit firm Apollo Global Management, asset manager BlackRock, private-equity firm Blackstone, asset manager Brookfield, Goldman Sachs, and private-equity firm KKR — aiming to mobilize more than $500 billion in outside capital for AI infrastructure over time [8]. When asked directly whether NVIDIA’s own money was behind that capital, CEO Jensen Huang said no, describing it as third-party, independent capital [9]. NVIDIA’s stock still fell 2.9% the day the deal was announced, an unusual reaction to a $500 billion headline [8], [10].

What came out afterward in writing complicates that clean answer. NVIDIA disclosed that on some projects it may provide “residual-value support” covering up to 25% of an individual opportunity, meaning NVIDIA itself could still absorb part of the loss if financed equipment ends up worth less than expected — a potential exposure as large as $125 billion if fully used, though not a debt NVIDIA owes today [9], [11]. Credit markets are already pricing that added risk: CoreWeave, a company that rents out NVIDIA chips by the hour, closed a $2.6 billion loan on its own GPU hardware that same Monday, August 10, at a full percentage point wider than a similar loan it closed in May, with a five-year maturity that runs longer than the roughly three-year average length of its own customer contracts [11]. The deal still got done and was oversubscribed, so capital has not dried up — it has simply gotten more expensive and more risk-aware [11].

We are treating this as the moment the AI trade started trading on credit-market terms as much as on earnings. We are watching how much of NVIDIA’s 25% support cap actually gets used deal by deal, and whether loan spreads on AI infrastructure keep widening the way CoreWeave’s did this week, because that is a better real-time read on risk appetite than any single company’s own growth forecast.

One day’s earnings-fueled rally in AI infrastructure stocks doesn’t offset the falling price of AI itself

The credit market’s caution about CoreWeave’s financing, described above, sat alongside a very different reaction in CoreWeave’s stock the same week — a reminder that credit investors and equity investors can look at the same company and price two different risks.

CoreWeave jumped more than 19% and Nebius, a cloud-computing company, rose 23% on Wednesday, August 12, after both posted earnings and issued upbeat forecasts, with Super Micro Computer, a server maker, gaining more than 13% the same day [12]. CoreWeave’s contracted revenue backlog rose to $104.2 billion, and Wall Street research firm MoffettNathanson said pricing trends suggest project returns are strengthening even as the underlying cost per unit rises [13].

That demand is real, but it is a different question from what is happening one layer up, at the AI models themselves. SpaceX is pitching its Grok 4.6 model as a cheap, near-frontier alternative [14], and OpenAI has been cutting prices on its smaller models as businesses push back on their AI bills [15]. Compute providers can stay fully booked even as the price of the intelligence running on that compute keeps falling toward zero — which is a fine outcome for infrastructure owners, but a real problem for anyone betting on durable pricing power at the model layer.

We are not changing anything in client portfolios based on one earnings-fueled trading day for a handful of AI infrastructure companies. The model-layer pricing pressure is a separate, slower-moving question, and it will not be resolved by a single day’s strong reaction in the stocks of the companies renting out the hardware underneath it.

An unconfirmed report about a sovereign fund’s intentions, not a disclosed purchase, moved Korean chip stocks

Samsung Electronics closed Wednesday, August 12’s Seoul trading session up about 6.7%, and SK Hynix, the South Korean memory-chip maker, gained about 5.5% the same day, after South Korean outlet Asia Business Daily reported that Temasek, Singapore’s state investment fund, was considering new direct investments in both companies and had contacted the Korean government about timing [13].

Temasek’s own spokesperson complicated that story the same day, saying the fund had not asked Seoul about investment timing and that it first invested in both Samsung and SK Hynix more than two years ago, pointing to other long-held Korean positions including Celltrion, Hyundai Motor, and POSCO [13], [16]. No new purchase has been confirmed. A long-term investor adding to a position it has held for years is a materially different signal than a state fund making its first-ever move into a sector, and the more dramatic version of the story is the one that remains unconfirmed.

This is the pattern worth remembering more than the stock move itself: a single-sourced report about an investor’s intentions can move a market before anyone has verified it, and the company named in the report doesn’t always confirm the version that moved the price. We are not chasing rallies built on that kind of report. A public disclosure of an actual position change would be worth revisiting — an unconfirmed one is not.